Humanize PPT
Use this skill when a user wants to turn raw material, notes, voice transcripts, documents, links, or old PPTs into a presentation-ready outline and per-page media decisions before delegating rendering to a downstream skill.
Positioning
Humanize PPT is a presentation system, born for the talk: an Outline Director (AST audience-state-transfer — every page turn moves the audience forward), a Per-Page Visual-Enhancement Director (real image / SVG diagram / Remotion video), a Production Brief Orchestrator, a Presentation Checkup Runner (演讲体检; formerly the QA loop, CLI flag still --qa-from), and a Presenter-Mode hand-off. The motivation: HTML-PPT template skills are great at concept display but blow a simple idea into a dozen pretty pages, while a real 90-minute talk is ~30 — the pretty shell outruns the content density. Humanize closes that gap: it keeps the beauty (rendered natively by the downstream template skill) and makes it presentable — a line you can stand up and deliver. Downstream template skills own "renders beautifully"; Humanize owns "it's a talk, and someone checked it."
The presentation checkup in one sentence: it does not grade beauty, it grades the outline. It compares every rendered page against its outline page, pulls out the pages that can only be looked at but not spoken from, and keeps going until every page is one the speaker can stand up and present. A failed page, in plain words: a page that holds only a few words and never finishes its point, or a page that fails the audience state transfer it promised (the listener walks out of that page in the same state they walked in). Such a page should not exist; the checkup pulls it out and generates fix instructions.
Humanize is broadly compatible with any downstream skill that can output an HTML PPT: the brief is plain markdown + JSON, anything can read it. The verified, stable recommendations are: Chinese → guizang-ppt-skill; English → frontend-slides / beautiful-html-templates. Other downstreams are hot-pluggable; support levels live in registry/renderer_registry.json and are updated only on real results.
It runs before downstream PPT / HTML slide skills and around the post-render presentation checkup. It owns the AST contract, the per-page media decision (does this page need a photo, a system diagram, a 10-second process clip, nothing?), the production brief that the next agent consumes, and the checkup pass on the rendered HTML. It does not own the rendered HTML itself.
There are two human review gates before rendering, both Humanize-owned zero-dependency working drafts (never decks): the outline preview (--preview-outline / scripts/preview_outline_html.py, the audience state-transfer map) and, since v0.9, the style gallery (--style-gallery). The style gallery is the cover-style gate: before the outline, Humanize emits ≥4 cover candidates for the renderer, writes one cover-only render command per candidate for the downstream skill to render, and stitches style_gallery.html to pick from. Humanize emits only the spec and the commands; the covers are rendered downstream. Picking a cover yields a re-injection command that carries the chosen style into the normal flow. See references/style-gallery-spec.md.
V0.6.4 is the single entrypoint for this loop. The user calls Humanize PPT once for the brief, hands the brief to a downstream skill for native rendering, then calls Humanize PPT again with --qa-from <rendered.html> to run the 3-iteration presentation checkup. Each iteration writes qa_report.md (findings), fix_prompt.md (downstream-skill-actionable corrections), and qa_iteration.json (round state). After 3 rounds with remaining findings, status flips to needs-human.
Humanize PPT never copies a downstream skill's template, never injects custom sections into it, and never post-processes the rendered HTML. When the downstream skill updates, Humanize PPT needs zero changes. This is the contract — see references/guizang-production-brief-orchestrator.md for the full brief specification.
For public positioning, describe Humanize PPT as a brief orchestrator that pairs with native downstream renderers. Do not frame it as a renderer itself, and do not present it as a "router" that picks the best visual style for the user — that decision lives in the brief, the downstream skill's own templates, and the human's review. When a user only wants a pretty template page, that is a rendering-skill job, not a Humanize job: state the choice, not a prohibition.
AST theory
AST means Audience-State-Transfer.
- Audience: who is listening, what they know, what they resist, and why they would keep listening.
- State: the audience state before and after the deck, plus the core tension that blocks the transition.
- Transfer: the slide-by-slide path that moves the audience from initial state to desired state.
Core sentence:
PPT is not an information container. PPT is an audience state-transfer artifact.
Required output contract
For every Humanize PPT run, produce:
deck_brief.md — audience, goal, tension, success criteria.ast_outline.md — AST map and narrative arc.slide_plan.json — slide-by-slide plan, with per-page media: {image, diagram, video} decision and layout_hint.speaker_intent.md — what the speaker should do on each slide. Downstream skills consume this as the source for their native speaker notes and presenter shell.asset_manifest.md — Humanize's per-page material decisions: which page needs which kind of asset (image / diagram / video) and for what purpose.video_slots.json — optional Remotion / HyperFrames / native video insertion plan.style_brief.md — visual principle for downstream production.renderer_registry.json — renderer capability snapshot for this run.router_plan.json — selected primary renderer and staged route plan.commands/*.md — bounded instructions for each downstream specialist agent.run_manifest.json — final file inventory, route status, and QA status.<renderer>-production-prompt.md — the production brief that the next agent consumes. v0.6.4 emits guizang-production-prompt.md, frontend-slides-production-prompt.md, or beautiful-html-templates-production-prompt.md depending on language and route. This file references the downstream skill's own SKILL.md and is the only thing the next agent needs to read.outputs/qa/qa_report.md — first-pass QA gate (brief mode) or per-iteration QA findings (QA mode).
QA mode (post-render) additionally produces per iteration:
outputs/qa/fix_prompt.md — downstream-skill-actionable fix instructions.outputs/qa/qa_iteration.json — round number, status (iterate / pass / needs-human), unresolved findings, history.
Style-gallery mode (--style-gallery, pre-outline gate) instead produces and stops:
style_gallery.html — zero-dependency picker stitching ≥4 candidate covers.style_gallery_plan.json — per-candidate id, cli, command_file, cover paths, re-injection command.commands/style-gallery/<id>.md — one cover-only render command per candidate (downstream renders only S01 → outputs/style-gallery/<id>/cover.{html,png}).
Recommended OPC workflow (v0.6.4)
O — Outline + Per-Page Media Direction
Humanize PPT: raw material → AST outline + per-page media decision
(deck_brief.md, ast_outline.md, slide_plan.json, speaker_intent.md,
asset_manifest.md, video_slots.json, style_brief.md)
P — Native Renderer Invocation (100% downstream)
zh → guizang-ppt-skill (Style A or B, native; recommended)
en → frontend-slides / beautiful-html-templates (native; recommended)
other HTML-PPT skills → hot-pluggable, same brief contract
Humanize emits the production prompt and stops. The downstream
skill renders the deck. Humanize does NOT copy templates, does
NOT inject SLIDES_HERE / [必填] replacements, does NOT add
postMessage bridges to the rendered HTML.
Q — Presentation Checkup (演讲体检) on the rendered HTML
Humanize --qa-from <rendered.html> reads the output of P,
compares pages against the outline, scans for failure modes
(references/qa-failure-modes.md), writes qa_report.md and
fix_prompt.md, tracks iteration in qa_iteration.json.
Cap: 3 rounds. After cap with remaining findings, status
flips to needs-human.
C — Complete / Control
Downstream skill native speaker notes + presenter shell + deploy
(Humanize does not own these in v0.6.4 — the brief tells the
next agent to produce them in the downstream skill's own format)
Rules
v0.6.4 invariants (these are the hard rules; if you break any, you're off the v0.6.4 boundary)
- Humanize is brief-only. It writes
<renderer>-production-prompt.md and stops. It does not open, copy, or post-process the downstream skill's template. When the downstream skill updates its template, animation markers, or validator, Humanize needs zero changes.
- Downstream renderers are 100% native. The next agent follows the downstream skill's own
SKILL.md. The brief tells the next agent which skill to load, which Style (A/B) to use, which layouts to pick from, and which QA gates must pass — but it does not carry template internals.
- The presentation checkup caps at 3 iterations. Round 4 with remaining fail findings is
needs-human. The loop does not spin forever; it hands the decision back to a human.
- Speaker notes and presenter shell — Humanize produces a baseline
presenter-shell.html; downstream owns the full stage. Humanize owns the semantic source (speaker_intent.md) and now also writes outputs/presenter/presenter-shell.html directly from slide_plan.json + speaker_intent.md (usable standalone, even before the downstream deck exists). The downstream skill produces the native speaker notes and fuller presenter console. Humanize does not inject postMessage bridges or ?slide= URL parameters into the rendered HTML.
- The brief is the only thing the next agent reads. It is a complete contract: deck metadata, per-page media decisions, style files, hard rules, known-good checkpoint, and per-style QA gates. The next agent does not need to re-derive intent from raw source material.
Working rules
- Do not let slide renderers consume raw material directly when Humanize PPT can first produce the AST contract.
- Keep the downstream skill as the owner of the full stage view; Humanize's
presenter-shell.html is a functional baseline, not a replacement for native consoles.
- Absorb AI-writing cleanup principles from humanizer tools, but do not reduce Humanize PPT to text polishing.
- Prefer a small verified workflow over a broad unverified promise.
- For public Skill releases, create/push the repo, install from GitHub locally, run one safe full sample, verify the brief + presentation checkup on the verified known-good checkpoint (
examples/03-codex-guizang-native-ink-classic/), and only then polish README details.
- For Agent Teams development, emit
router_plan.json, run_manifest.json, bounded commands/*.md, and the per-renderer production prompt before wiring real downstream Skills.
- For WorkBuddy/CodeBuddy team upload packages, do not package demo or rendered HTML outputs as the team zip. The upload zip must mirror a team-plugin structure like
trading-team: root-level .codebuddy-plugin/plugin.json, agents/, skills/, rules/, and setting.json (plus optional avatars/, .workbuddy-plugin/, README.md, settings.json). The rules/ directory should include a scenario rule file such as rules/<plugin-name>_rules.md with frontmatter (description, alwaysApply, enabled, updatedAt, provider) and a <system_reminder> block describing available agents, skills, SOP, and usage requirements. Verify with unzip -l that the root is not index.html/assets/screenshots/source and is not folder-wrapped unless the target uploader explicitly requires a wrapper directory.
- Do not treat HyperFrames/Remotion videos as a single embedded player that replaces PPT content. For Humanize PPT deliverables, video tools are material producers: transitions, explainer clips, before/after comparisons, talking-material inserts, social previews, and fallback stills that fill specific slide needs. The
media.video decision per page (see slide_plan.json schema) tells the downstream skill which pages want a Remotion clip, for what purpose, and at what duration.
Renderer-specific guidance (kept for history; the boundary itself is the invariants above)
- For Chinese PPT production, the recommended stable path is
Humanize PPT → guizang-ppt-skill native → Humanize --qa-from → downstream presenter/deploy. Guizang's own material QA and Swiss validator run inside the downstream skill. The presentation checkup in Humanize is a second-pair-of-eyes pass, not a replacement.
- For English PPT production, the recommended path is
Humanize PPT → frontend-slides or beautiful-html-templates (native) → Humanize --qa-from → downstream deploy. The downstream skill owns its own template selection, preview gallery, and selected-template full deck. Humanize does not imitate them. v1.0 support levels (see registry/renderer_registry.json): both beautiful-html-templates and frontend-slides are now full — brief exit works, full presentation checkup ran on real decks, and 5 renderer-specific failure-mode rules added (horizontal overflow, low contrast, hyphenation noise, font contract, image alt). Screenshot review is still part of the checkup methodology for cases the static scan can't reach.
- The verified Style A checkpoint at
examples/03-codex-guizang-native-ink-classic/ is a read-only visual reference. If the presentation checkup ever fails against it (test_known_good_style_a_passes_all_style_a_gates), the fixture or the live Guizang skill has drifted — do not weaken the checkup to make the test pass. The same applies to the English checked-up deck at docs/showcase/hermes-agent-mastery/en/ppt/ (test_english_showcase_deck_passes_presentation_checkup).
Operational references
references/guizang-production-brief-orchestrator.md — v0.6.4 canonical brief specification. The human + agent-facing contract for what <renderer>-production-prompt.md must contain and what it must not contain.SPEC.md — engine technical specification: boundary, CLI surface (mode-check order), data flow, output contract, the v0.9 style gallery, the presentation checkup, the per-page media model, and the renderer registry. The authoritative "what the engine does and guarantees" reference.references/qa-failure-modes.md (+ English mirror references/qa-failure-modes.en.md) — failure mode catalog for the presentation checkup (演讲体检): a renderer-agnostic failure-class layer plus guizang-specific modes, each with what the audience would see. Human-readable; the code-side source of truth is FAILURE_MODES in scripts/humanize_ppt_v2.py. Includes the WebGL-hero static-screenshot capture trap as a "static scan can't catch yet" class.references/style-gallery-spec.md — v0.9 spec for the --style-gallery cover-style gate: candidates, cover-only commands, the zero-dependency picker, re-injection, and the WebGL screenshot warning.scripts/preview_outline_html.py — outline preview: renders the audience state-transfer map (one zero-dependency single-file HTML; per-slide enter-state → intent → leave-state rows plus a state-arc summary) from slide_plan.json. Real sample: examples/04-preview-outline-ai-tool-update/.scripts/record_demo_gif.py — records the style gallery + outline preview (the two zero-dependency working drafts) into one demo GIF (requires playwright + ffmpeg). The gallery covers are downstream-rendered; --covers-dir overlays real covers before recording.docs/versions/v0.9.0-style-gallery.md — v0.9 release notes: the cover-style gallery gate, the WebGL static-screenshot failure class, the English failure-mode mirror, SPEC.md, and the README/GIF-slot work.docs/versions/v0.8.0-presentation-checkup.md — v0.8.0 release notes: why the QA loop was renamed to presentation checkup, the hot-pluggable route framing, the plain-language usage rewrite, and the verified English checkup run.docs/versions/v0.7.0-render-qa-inspector.md — v0.7.0 release notes: why the positioning moved to render-QA inspector, English-path support levels, and the outline preview artifact.references/agent-teams-public-preview.md — Agent Teams architecture, specialist-agent command protocol, public preview release loop, and README split convention. (Historical; v0.6.4 collapses the Agent Teams model into a brief + QA loop.)
Local demo
The recommended stable entrypoint is scripts/humanize_ppt.py. Versioned scripts remain available for compatibility.
Brief mode (v0.6.4 default — writes a Guizang production brief, no HTML):
python3 scripts/humanize_ppt.py \
--source examples/01-ai-tool-update/source.md \
--out .humanize-ppt-runs/ai-tool-update-v0.6.4 \
--title "AI 工具更新，不只是功能清单" \
--renderer guizang \
--guizang-style A
The next agent reads guizang-production-prompt.md and renders natively via guizang-ppt-skill. Once the deck is rendered, run the presentation checkup:
python3 scripts/humanize_ppt.py \
--qa-from .humanize-ppt-runs/ai-tool-update-v0.6.4/rendered/index.html \
--out .humanize-ppt-runs/ai-tool-update-v0.6.4 \
--renderer guizang \
--guizang-style A \
--max-qa-iterations 3
English paths use the same shape with --renderer beautiful-html-templates or --renderer frontend-slides, which write beautiful-html-templates-production-prompt.md or frontend-slides-production-prompt.md respectively. (v0.9: both beautiful-html-templates and frontend-slides are support_level: brief+qa-verified — beautiful after the 2026-06-13 real-deck checkup, frontend-slides after a real-deck checkup on 2026-06-17. The v0.9 style-gallery gate itself changed no support level — those move only on real rendered output.)
Outline preview (audience state-transfer map from an existing slide_plan.json, zero-dependency single-file HTML):
python3 scripts/preview_outline_html.py \
--slide-plan .humanize-ppt-runs/ai-tool-update-v0.6.4/slide_plan.json \
--out .humanize-ppt-runs/ai-tool-update-v0.6.4/preview-outline.html \
--title "AI 工具更新，不只是功能清单"
The legacy V0.2-compatible entrypoint remains available for compatibility with earlier agents:
python3 scripts/humanize_ppt_v2.py \
--source examples/01-ai-tool-update/source.md \
--out .humanize-ppt-runs/ai-tool-update-v0.2 \
--title "AI 工具更新，不只是功能清单" \
--renderer auto
Legacy V0.1 demo remains available:
python3 scripts/humanize_ppt_v1.py \
--source examples/01-ai-tool-update/source.md \
--out .humanize-ppt-runs/ai-tool-update \
--title "AI 工具更新，不只是功能清单"
